Dunhuang Mural Line Drawing Based on Bi-Dexined Network and Adaptive Weight Learning.

Baokai Liu,Shiqiang Du,Jiacheng Li,Jianhua Wang,Wenjie Liu

DOI: https://doi.org/10.1007/978-3-031-18907-4_22

2022-01-01

Abstract:Dunhuang murals are an excellent cultural heritage, a masterpiece of Chinese painting, and a treasure of Buddhist art. A large number of murals and sculptures have gone through thousands of years and are of high artistic value. Digital line drawings of the Dun- huang Murals can not only show the beauty of the line art of murals, but also guide the restoration of murals. It belongs to the direction of image edge detection in computer vision. The purpose of edge detection is to quickly and accurately locate and extract image edge feature information. Although some traditional detection algorithms and methods based on deep learning have made some progress, they have not achieved ideal results in generating mural datasets. Compared to existing methods, we propose a novel edge detection architecture Bi-Dexined network. Firstly, the adaptive weight in this method can well balance the weight influence of the fusion of different levels of feature maps on the final prediction result. Secondly, The upsampling path can extract deeper semantic information in the network. After testing on several different edge detection and mural datasets, our method can generate clearer and more reasonable edge maps than other methods.